Jamie

Data Scientist in Video Games & Entertainment industry

[Jamie @Randum_00]

If there's one thing you'd think large language models are supposed to be good at, it's language. https://t.co/hZnIbnHPxZ

[Jamie @Randum_00]

@thsottiaux UI/UX feature for agent - UI collaboration.

Please add push notifications for when the agent needs context or human input. This could be as simple as codex sending a notification (incl on iOS app) when a run is done OR better as a hook mid-run.

[Jamie @Randum_00]

@adonis_singh Generative UI powered by GPT on Cerebras.

[Jamie @Randum_00]

@gdb @gdb when will it narrate the book or pdf? I want this feature..

[Jamie @Randum_00]

@joseph_h_garvin @METR_Evals I susoect METR didnt tell the model that this behaviour was “cheating”.

e.g. if I tell you to solve a hard math problem, and you find the problem on the internet and share the solution, technically I’ve delivered the result, but METR may call this cheating.

[Jamie @Randum_00]

@hwchase17 @TheEthanDing I've written SQL my whole career, and I don't write SQL any more. Everything I do is in natural language due to these changes.

[Jamie @Randum_00]

@hwchase17 @TheEthanDing @hwchase17 - the models.. They're now more reliable.
1. Connect the model to the code base that created the db
2. Connect the model to the docs for what columns mean
3. Models can list schema and retry.

These all make it now viable (and working) where it wasn't before.

[Jamie @Randum_00]

@embirico @OpenAI @embirico a Jupyter notebook extension that autonomously completes analysis tasks based on a prompt input. Around 3 weeks since first commit). Codex for everything (except some UI components).
* multi-provider
* permissions system
* Auth & user security
& more.

[Jamie @Randum_00]

Give agents access & capabilities without leaking API keys.

https://t.co/zwzsE7Ub1M

[Jamie @Randum_00]

@TheMingjie I'd appreciate a key loaded with $500-$1k for a side project I'm working on - it's a notebook assistant for working with Research & Data Scientists in Jupyter notebooks.

[Jamie @Randum_00]

@venturetwins @venturetwins

Difficulty: easy

https://t.co/91cUp76xTk

[Jamie @Randum_00]

@goodside @goodside

[Jamie @Randum_00]

@goodside Clapping is not required. https://t.co/RujikIrwX5

[Jamie @Randum_00]

@venturetwins @venturetwins this worked fine. This worked fine and ChatGPT responded with David Mayer - https://t.co/o1HQqjPWaU

[Jamie @Randum_00]

@simonw I think it’s more likely that as safety/usability features are built-in, helpfulness/performance falls.
In addition, there’s likely a trade off with turbo between making cheaper compute and faster inference speed vs. Accuracy and performance.

[Jamie @Randum_00]

@EMostaque Companies will have a $ budget or compute budget, not a “performance” budget.

The demand for GPUs should in the most part hold if the budget projections are accurate.

In other words, if such a 10% improvement is possible, just train a bigger model instead for longer.

[Jamie @Randum_00]

GPT in Jupyter notebook.
https://t.co/z4m9CZXsq5

[Jamie @Randum_00]

@AlphaSignalAI Totally agree..
Too convenient ->
https://t.co/z4m9CZXsq5

[Jamie @Randum_00]

@RadissonBlu Disappointed that we had no clean water access during our stay in Mysore, India. Also, the hotel couldn’t offer a satisfactory fix.

Look at the colour (while running a bath to bathe our 1 year old).

[Sean J. Taylor @seanjtaylor]

This is a nice Bayesian Optimization blog post by some scientists at Apple (including a discussion of @DrewDim's excellent paper on shrinkage estimators). Notable to see folks at Apple sharing this much information about their work. https://t.co/UJyyC2ouB8